Reproduce “Local-GS: Accelerating 3D Gaussian Splatting via Tile-Local Warp Coherence” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Yang Luo, Yan Gong, Yongsheng Gao, Jie Zhao, Xinyu Zhang, Huaping Liu
Local-GS targets interactive 3D rendering by reorganizing Gaussian-splatting work around GPU warps rather than only tiles. It hoists tile-shared Gaussian calculations, culls warps outside visible footprints, and uses warp-coherent blending. On RTX 4080, FlashGS plus Local-GS reaches 1601.77 FPS versus 206.37 for vanilla 3DGS on Deep Blending, a reported 7.76× speedup, with matching quality metrics in that comparison.
Key insights
- Optimizing 3DGS rendering efficiency requires shifting focus from coarse-grained primitive culling to fine-grained GPU execution coherence by aligning processing with SIMT boundaries.
- Re-parameterizing Gaussian alpha calculation into a tile-local polynomial form drastically reduces redundant arithmetic and improves numerical stability by mitigating floating-point cancellation errors.
- Directly mitigating warp divergence and redundant computation at the kernel level, through techniques like warp-level culling and branch-free blending, significantly boosts GPU throughput without compromising visual quality.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.